[FEA] CUDA DLLs from Torch GPU

[gmarkall]

From @fbordignon in numba/numba#10074:

Hi guys, I would like to run an Idea by you.
We are deploying numba within an app that already has torch with packaged cuda dlls

Would you consider merging a change to add this option for CUDA installation?

The idea is to change cuda_paths.py by adding the torch-cuda option and auxiliar functions to find the dlls inside the site-packages folder.

tyia

[gmarkall]

I just realised this should really become a question for the CUDA Python Path Finder rather than Numba-CUDA's implementation - @rwgk what do you think of this suggested change? (Should this issue move to https://github.com/NVIDIA/cuda-python?)

[leofang]

We already handle the CUDA wheel lookup/loading that PyTorch and all other projects need, so feel free to close this issue! 🙂
https://github.com/NVIDIA/cuda-python/blob/main/cuda_bindings/cuda/bindings/_path_finder/README.md

[leofang]

(For the record there are some loose ends that @rwgk will address to make it useable by end users, the tracking epic is here: NVIDIA/cuda-python#451)

[gmarkall]

Thanks @leofang - Numba-CUDA still needs to make use of the path finder, so I'll leave this open until that's done.

[leofang]

Numba-CUDA still needs to make use of the path finder

I think this is already tracked once we complete #128 (and its child issues).